CS498PS - Lab 3: Artificial Reverberation and Room Simulation¶
Ching-Hua Yu (cyu17)¶
In this lab we will get some experience in simulating rooms digitally. We will design some simple reverbs using small filters and then work towards a more principled room simulator. You can use whichever sound you like for the following examples, usually a sound with sharp onsets is good for such experiments. One good example is the drum loop [ https://drive.google.com/uc?export=download&id=16mqI3xFN00AK5bP5k1VYJiEfH7ghHHSf ]
Part 1. Designing simple reverb using filters¶
The simplest forms of reverbs can be designed using simple delays, comb and allpass filters. We will design one of each to get started.
a) A room with a single wall. For this reverb we will assume that we have a room that only has one reflective wall as shown below. Design a filter that simulates what the microphone will record. figure out the right filter in the case where $d_i = [0.1, 1]$ meters, and in the case where $d_i = [1, 10]$ meters. Let me know what the differences between these sounds are. To keep things simple, round the delays to an integer number and make up an approximate gain loss due to propagation and the wall bounce.
The main difference between d=[0.1,1] and d=[1,10] is that the echo in the later case (long distance) is much clear. The audio in the first case (short distance) is just very similar to the original sound. I think this also related to the choice of the input sound.
b) Now we will use a room geometry as shown below. Design the proper filter again, for $d_i = [0.1, 1]$ meters and then for $d_i = [1, 10]$ meters. Comment on the audible differences between these two settings.
The impulse response from two walls is much stronger than that of one wall in case of the long distance. However, for the short, the echo is not very clear except small change in the sound wave.
c) Since for small values of $d$ the output above gets its spectrum altered, we will use an allpass filter to make a more natural sounding echo pattern. Use the formulation shown in the lecture slides and compare that filter’s frequency response with the response from the filter in the section b) above.
The result of allpass filter is similar to that of comb filter in (b), but the echo has some more cancelation with the original sound and looks more natural.
Part 2. Schroeder reverberators¶
Now we will combine multiple filters to make a more serious sounding reverb. Implement the structure shown below.
Try to find the necessary parameters for the above structure that makes it sound as good as possible (not too busy, not too subtle). Hint: There is no right answer! Try a few things, see what sounds best and explain why you used the parameters you decided to use. As you come across some bad sounding cases, describe what the problem was.
I used different distances as the parameter in comb filter. Not sure if that makes sense in the real case, but it seems the two allpass filters performed well in difussion and controled the output.
Part 3. Applying a real room response¶
To apply a more realistic filter we need to use a real room response. Download the St. Andrew’s church room impulse response (Stereo) from:
[ https://openairlib.net/?page_id=683 ]
Make sure that the sound you will convolve with it will be at the same sampling rate. This impulse response captures that church’s RIR using two channels. To synthesize a sound in that room you need to convolve it with the impulse response. If your test sound is a single-channel recording, you can simply convolve it with each of the two impulse responses and then use the resulting outputs as the left and right channel. You will notice that using the scipy.signal.lfilter command for this operation can take very long to compute. For faster convolutions we need to perform this filtering in the frequency domain (using the fast convolution algorithm that we mentioned in the filtering lecture). The function numpy.convolve automatically does that (if you use an up to date version), otherwise you can directly use scipy.signal.fftconvolve. Verify that the resulting output sounds like it has been placed inside a church.
Part 4 (extra credit). Measuring a room response.¶
Measuring a room response is crucial when trying to emulate a real room. In this exercise we will make a simple room response measurement. We will use the technique shown in the lecture slides.
a) Generate a chirp to use as our room excitation signal.
b) Go to a (preferably quiet) room that you want to measure, play the chirp from a speaker and record the outcome (the easiest way to do that is to save this sequence as a wave file and play it while you record a new sound file). You can instead use the real-time I/O code from Lab 0. In doing so you will obtain the sequence $y[t]$ from slide 36.
c) Now you will need to perform the actual deconvolution. Take $x[t]$ and $y[t]$, compute their DFTs to get $X[ω]$ and $Y[ω]$, and estimate $H[ω]$ (remember to zero pad the inputs since the output of the convolution will be about twice as long as the inputs). Take the inverse DFT of $H$ to get the time domain representation of the room impulse response. Does it look like one?
d) Convolve the resulting response with the test sound above (make sure you use the same sample rates). Does it sound reverberated? This won’t be perfect. What sounds “wrong” with it? Can you speculate why?
